{ "schema_version": 1, "builder_plan": { "scenario": "A self-hosted assistant is too expensive because its model uses most of the GPU memory.", "change": "Try the paper's supported lower-precision weights on a copy of your serving setup.", "baseline": "The same model at its original precision, on the same GPU.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Peak memory, task pass rate, latency, and total serving cost at the same context and load.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2610.30480", "title": "FIGFLEX: A Mixed-Precision FP-INT GEMM Accelerator Co-Designed with Low-Bit LLM Quantization", "summary": "FIGFLEX replaces GPU dequantize-to-FP16 tensor-core execution of weight-only quantized models with a custom bit-parallel FP-INT accelerator whose throughput scales as weight precision falls, paired with a mixed-precision quantizer that assigns per-group bit-widths. The approach requires non-commercial accelerator hardware (an ASIC/FPGA-class design), so adopting the performance gains is impossible on a 24 GB GPU or API-based setup; only the quantizer is potentially portable, at the cost of added quantization complexity. Without the custom datapath, the claimed throughput scaling does not materialize, and the accuracy results depend on the co-designed scaling scheme that has been validated only on LLaMA-2-7B WikiText-2.", "rationale": "The core contribution is custom accelerator hardware unavailable under the reader's constraints, so the work is worth monitoring but not adoptable in production.", "technique": "Precision-flexible FP-INT GEMM accelerator with co-designed quantization", "family": "quantizacao", "recommendation": "observar", "published": "2026-10-21", "as_of": "2026-10-24", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.30480", "independent_implementations": 0, "total_implementations": 1, "stars": 4, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "FIGFLEX replaces GPU dequantize-to-FP16 tensor-core execution of weight-only quantized models with a custom bit-parallel FP-INT accelerator whose throughput scales as weight precision falls, paired with a mixed-precision quantizer that assigns per-group bit-widths. The approach requires non-commercial accelerator hardware (an ASIC/FPGA-class design), so adopting the performance gains is impossible on a 24 GB GPU or API-based setup; only the quantizer is potentially portable, at the cost of added quantization complexity. Without the custom datapath, the claimed throughput scaling does not materialize, and the accuracy results depend on the co-designed scaling scheme that has been validated only on LLaMA-2-7B WikiText-2.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.30480", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "On LLaMA-2-7B, 5.85 and 8.77 WikiText-2 perplexity at 3- and 2-bit average precision, lower than GPTQ, AWQ, SliM-LLM, AMQ, and QuIP#; 1.56x and 1.42x higher throughput per area than AxCore and FIGLUT on 3-bit mixed-precision GEMMs.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.30480", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }